VideoJev vs Jev 1.13: Same Idea, Two Different Jobs

The comparison looks obvious until you try to write it. VideoJev is an open-weights research release from NJU's MCG lab that scores temporal intervals and answer options against video evidence, in two checkpoints, VideoJev-4B and VideoJev-35B-A3B. Jev 1.13 is TypeSafe AI's closed, hosted decision model — the model that named the category both of them live in. VideoJev's own paper cites Jev as the origin of the paradigm it extends, which makes this look like a head-to-head. It mostly is not. Jev 1.13 takes text state and returns typed decisions. VideoJev takes video and returns scored intervals or a chosen option. They overlap on the idea and almost nowhere on the input, so the useful question is not "which is better" but "which one is even eligible for the job you have."

Both arrived quietly and recently, in different ways. Jev 1.13 launched on 2026-09-15 — TypeSafe's first public model after two years in stealth — and has been callable through OrcaRouter's own catalogue at typesafe/jev-1.13 since 2026-09-24. VideoJev has no launch at all: the arXiv preprint landed 2026-10-13, the two Apache-2.0 checkpoints went up on Hugging Face on 2026-10-10, and the repository was still receiving README fixes on 2026-10-14. Neither is a launch story today. What makes the pairing worth reading is that they are the same architectural bet applied to two different data types, and the difference in how each is delivered tells you most of what you need to decide.

One line each, because the summaries diverge fast

Jev 1.13 is a hosted endpoint. You POST a block of state and a set of named questions to /v1/systemone and get back one typed answer per question — noul (a true/false with a calibrated probability), choice (one label from up to 255 options), or score (a level on an ordered 2-to-10 scale). Nothing is generated, so there are no output tokens and TypeSafe bills none; input is $0.042 per million tokens. Its architecture, parameter count and training compute are undisclosed. Its input is text only — strings, JSON, arrays — with no image, audio or video path. You do not run it; every call is metered.

VideoJev is the opposite arrangement. It is a checkpoint you download and serve yourself, under Apache-2.0, built by fine-tuning an existing multimodal LLM: VideoJev-4B on Qwen3.5-4B, VideoJev-35B-A3B on Qwen3.6-35B-A3B. The authors strip the final MLP and the language-model output head and replace them with a candidate scorer, so the model never writes an answer — it scores candidate intervals and candidate answers directly. Two inference modes organise it: Cached inference encodes the video once and branches each new query off that prefix, and Joint inference puts a group of questions behind one video prefix and reads them all out in a single forward pass.

The dimensions that actually separate them

• Input modality — video, at 2 fps with frames capped at 640×640 for VideoJev, vs text state only for Jev 1.13.

• What a "decision" is — a temporal interval in seconds plus an option index vs a typed label, level or probability.

• Delivery — self-hosted Apache-2.0 weights vs a metered hosted endpoint.

• Price — your own GPU hours vs $0.042 per million input tokens with output billed at zero.

• Cost of a second query on the same context — VideoJev's whole design problem, answered by prefix reuse; Jev re-sends state each call, since it has no video-sized prefix to amortise.

• Published detail — a nine-page preprint with per-benchmark tables and ablations vs no architecture, parameters or training disclosure.

• Where you can run it — anywhere you have a suitable GPU, vs TypeSafe's endpoint and the gateways that carry it, including typesafe/jev-1.13 on OrcaRouter.

Note what is not on that list: a shared benchmark. There is nothing both models have been scored on, because there is no task both can accept. Any table putting them side by side would be inventing a comparison.

Where the lineage is real, and where it stops

The intellectual debt is explicit rather than inferred. VideoJev's introduction frames its contribution as taking the System One premise — that a model can emit a structured judgment instead of generating tokens — and asking what changes when the thing being judged has a timeline. Jev's answer to "how does a model decide without writing?" is a scoring head over a declared option set. VideoJev's answer is the same scoring head pointed at timestamps as well as labels.

That extension is not cosmetic, and the paper is unusually specific about why. When the underlying context is a video, the expensive part is the visual encoding, and an agent that asks six questions about the same footage should not pay for it six times. Both of VideoJev's inference modes exist to fix that, and the numbers show what they buy: on ten grounding queries against one video, 0.39 s for VideoJev against 5.98 s for an Instruct baseline; on the QA side, ten queries in 1.67 s. Jev has no equivalent because a text state is cheap to resend — the problem Jev solves is decision latency per call, and the problem VideoJev solves is decision latency per call after the first.

The stopping point is modality. TypeSafe's own documentation is plain that System One input is text — "strings, JSON objects, and arrays of text" — and that images, audio and video are unsupported. So the two models are not substitutes on any task. If your state is a document, a chat log or a form submission, VideoJev cannot take it. If your question is "when in this footage does the woman in green eat pasta," Jev 1.13 cannot take it either.

The numbers, and who produced them

VideoJev's headline figures come from the authors' own harness and have not been independently reproduced. VideoJev-4B reaches 65.3 mean mIoU across three temporal-grounding benchmarks, 6.7 points above its Qwen3.5-4B backbone; VideoJev-35B-A3B reaches 67.7, 10.8 points above Qwen3.6-35B-A3B. On four video-QA benchmarks the two sit at 74.3 and 79.5 mean, within 0.6 points below and 0.7 points above their Thinking-mode backbones, at 129× and 100× the speed. The paper also runs a controlled comparison — an ordinary autoregressive Qwen3.5-4B fine-tuned on exactly the same data, same schedule — and VideoJev beats it by 1.3 points of grounding mIoU and 1.9 points of QA score. That control is the strongest evidence in the paper, because it isolates the readout shape from the training data.

Two caveats matter if you are planning to rely on it. The paper reports mean calibration error of 8.65 points at 4B and 8.13 at 35B-A3B, which means the scores are a good ranking signal and a poor probability — treat a 0.9 as "top of the list," not "90% likely." And the 35B checkpoint's headline configuration is FP32, per the project page, so the larger model's memory footprint is heavier than the parameter count suggests.

Jev 1.13's figures are vendor-published too, but they are the kind that can be checked by calling it: a fixed $0.042 per million input tokens, zero output cost, and 70–500 ms end-to-end response times. TypeSafe is also candid about which claims are unfalsifiable and which are not, noting that "no type errors" cannot be disproven with a counter-example because a value outside your declared set is not a value the model can emit.

If you need both, you have two procurement problems

This is where the difference in delivery stops being philosophical. VideoJev arrives as a repo: PyTorch 2.8, CUDA 12.8, FlashAttention 2, trust_remote_code=True, and a quickstart that assumes an H200 for the long-video settings. There is no vLLM or SGLang support yet, so it is not something you serve at throughput tomorrow. You are taking on an inference stack, not a vendor relationship.

Jev 1.13 arrives as an API key, and this is the one place OrcaRouter is directly relevant: the model is on our catalogue, so it sits on the same key as the generative half of whatever workflow you are building — one API for 200+ models, 0% markup (provider list price passed through, so a TypeSafe price change is live here the same day), automatic failover across providers, and a routing DSL if you want to compose several models into one call. VideoJev is not hosted here and this article will not pretend otherwise. Running it means downloading Apache-2.0 weights and standing up your own service; what we take off your plate is everything around it — the summariser that writes up the result, the cheaper classifier that gates whether the video call is worth making at all, the fallback when a provider wobbles.

The decision rule

The honest summary is that these two are a family resemblance, not a rivalry. Pick Jev 1.13 when the judgment is over text, the answer is one of a set you declare in advance, and you want someone else running the infrastructure and the calibration work. Pick VideoJev when the evidence is a video, the answer is a timestamp or a multiple-choice pick, and you have the GPUs and the appetite to run an unreplicated research release — with a hard look at your own footage first, since the paper documents its own failure mode (small-object counting, where both sizes miss a four-book answer) and yours will be different.

What would change the picture is a third checkpoint rather than a winner. Neither model is close to a general visual agent: Jev cannot see, and VideoJev can only answer the two question shapes it was trained on. The interesting news would be a VideoJev release with a text path, or a Jev release that accepts frames — at which point the two lines converge and this stops being two articles about two paradigms and becomes one article about a fight.

A release is also the moment to check what has gone wrong before: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.